1
Fourier Transform {#tutorial_py_fourier_transform}
7
In this section, we will learn
8
- To find the Fourier Transform of images using OpenCV
9
- To utilize the FFT functions available in Numpy
10
- Some applications of Fourier Transform
11
- We will see following functions : **cv2.dft()**, **cv2.idft()** etc
16
Fourier Transform is used to analyze the frequency characteristics of various filters. For images,
17
**2D Discrete Fourier Transform (DFT)** is used to find the frequency domain. A fast algorithm
18
called **Fast Fourier Transform (FFT)** is used for calculation of DFT. Details about these can be
19
found in any image processing or signal processing textbooks. Please see Additional Resources_
22
For a sinusoidal signal, \f$x(t) = A \sin(2 \pi ft)\f$, we can say \f$f\f$ is the frequency of signal, and
23
if its frequency domain is taken, we can see a spike at \f$f\f$. If signal is sampled to form a discrete
24
signal, we get the same frequency domain, but is periodic in the range \f$[- \pi, \pi]\f$ or \f$[0,2\pi]\f$
25
(or \f$[0,N]\f$ for N-point DFT). You can consider an image as a signal which is sampled in two
26
directions. So taking fourier transform in both X and Y directions gives you the frequency
27
representation of image.
29
More intuitively, for the sinusoidal signal, if the amplitude varies so fast in short time, you can
30
say it is a high frequency signal. If it varies slowly, it is a low frequency signal. You can extend
31
the same idea to images. Where does the amplitude varies drastically in images ? At the edge points,
32
or noises. So we can say, edges and noises are high frequency contents in an image. If there is no
33
much changes in amplitude, it is a low frequency component. ( Some links are added to
34
Additional Resources_ which explains frequency transform intuitively with examples).
36
Now we will see how to find the Fourier Transform.
38
Fourier Transform in Numpy
39
--------------------------
41
First we will see how to find Fourier Transform using Numpy. Numpy has an FFT package to do this.
42
**np.fft.fft2()** provides us the frequency transform which will be a complex array. Its first
43
argument is the input image, which is grayscale. Second argument is optional which decides the size
44
of output array. If it is greater than size of input image, input image is padded with zeros before
45
calculation of FFT. If it is less than input image, input image will be cropped. If no arguments
46
passed, Output array size will be same as input.
48
Now once you got the result, zero frequency component (DC component) will be at top left corner. If
49
you want to bring it to center, you need to shift the result by \f$\frac{N}{2}\f$ in both the
50
directions. This is simply done by the function, **np.fft.fftshift()**. (It is more easier to
51
analyze). Once you found the frequency transform, you can find the magnitude spectrum.
55
from matplotlib import pyplot as plt
57
img = cv2.imread('messi5.jpg',0)
59
fshift = np.fft.fftshift(f)
60
magnitude_spectrum = 20*np.log(np.abs(fshift))
62
plt.subplot(121),plt.imshow(img, cmap = 'gray')
63
plt.title('Input Image'), plt.xticks([]), plt.yticks([])
64
plt.subplot(122),plt.imshow(magnitude_spectrum, cmap = 'gray')
65
plt.title('Magnitude Spectrum'), plt.xticks([]), plt.yticks([])
68
Result look like below:
70
![image](images/fft1.jpg)
72
See, You can see more whiter region at the center showing low frequency content is more.
74
So you found the frequency transform Now you can do some operations in frequency domain, like high
75
pass filtering and reconstruct the image, ie find inverse DFT. For that you simply remove the low
76
frequencies by masking with a rectangular window of size 60x60. Then apply the inverse shift using
77
**np.fft.ifftshift()** so that DC component again come at the top-left corner. Then find inverse FFT
78
using **np.ifft2()** function. The result, again, will be a complex number. You can take its
81
rows, cols = img.shape
82
crow,ccol = rows/2 , cols/2
83
fshift[crow-30:crow+30, ccol-30:ccol+30] = 0
84
f_ishift = np.fft.ifftshift(fshift)
85
img_back = np.fft.ifft2(f_ishift)
86
img_back = np.abs(img_back)
88
plt.subplot(131),plt.imshow(img, cmap = 'gray')
89
plt.title('Input Image'), plt.xticks([]), plt.yticks([])
90
plt.subplot(132),plt.imshow(img_back, cmap = 'gray')
91
plt.title('Image after HPF'), plt.xticks([]), plt.yticks([])
92
plt.subplot(133),plt.imshow(img_back)
93
plt.title('Result in JET'), plt.xticks([]), plt.yticks([])
97
Result look like below:
99
![image](images/fft2.jpg)
101
The result shows High Pass Filtering is an edge detection operation. This is what we have seen in
102
Image Gradients chapter. This also shows that most of the image data is present in the Low frequency
103
region of the spectrum. Anyway we have seen how to find DFT, IDFT etc in Numpy. Now let's see how to
106
If you closely watch the result, especially the last image in JET color, you can see some artifacts
107
(One instance I have marked in red arrow). It shows some ripple like structures there, and it is
108
called **ringing effects**. It is caused by the rectangular window we used for masking. This mask is
109
converted to sinc shape which causes this problem. So rectangular windows is not used for filtering.
110
Better option is Gaussian Windows.
112
Fourier Transform in OpenCV
113
---------------------------
115
OpenCV provides the functions **cv2.dft()** and **cv2.idft()** for this. It returns the same result
116
as previous, but with two channels. First channel will have the real part of the result and second
117
channel will have the imaginary part of the result. The input image should be converted to
118
np.float32 first. We will see how to do it.
122
from matplotlib import pyplot as plt
124
img = cv2.imread('messi5.jpg',0)
126
dft = cv2.dft(np.float32(img),flags = cv2.DFT_COMPLEX_OUTPUT)
127
dft_shift = np.fft.fftshift(dft)
129
magnitude_spectrum = 20*np.log(cv2.magnitude(dft_shift[:,:,0],dft_shift[:,:,1]))
131
plt.subplot(121),plt.imshow(img, cmap = 'gray')
132
plt.title('Input Image'), plt.xticks([]), plt.yticks([])
133
plt.subplot(122),plt.imshow(magnitude_spectrum, cmap = 'gray')
134
plt.title('Magnitude Spectrum'), plt.xticks([]), plt.yticks([])
138
@note You can also use **cv2.cartToPolar()** which returns both magnitude and phase in a single shot
140
So, now we have to do inverse DFT. In previous session, we created a HPF, this time we will see how
141
to remove high frequency contents in the image, ie we apply LPF to image. It actually blurs the
142
image. For this, we create a mask first with high value (1) at low frequencies, ie we pass the LF
143
content, and 0 at HF region.
146
rows, cols = img.shape
147
crow,ccol = rows/2 , cols/2
149
# create a mask first, center square is 1, remaining all zeros
150
mask = np.zeros((rows,cols,2),np.uint8)
151
mask[crow-30:crow+30, ccol-30:ccol+30] = 1
153
# apply mask and inverse DFT
154
fshift = dft_shift*mask
155
f_ishift = np.fft.ifftshift(fshift)
156
img_back = cv2.idft(f_ishift)
157
img_back = cv2.magnitude(img_back[:,:,0],img_back[:,:,1])
159
plt.subplot(121),plt.imshow(img, cmap = 'gray')
160
plt.title('Input Image'), plt.xticks([]), plt.yticks([])
161
plt.subplot(122),plt.imshow(img_back, cmap = 'gray')
162
plt.title('Magnitude Spectrum'), plt.xticks([]), plt.yticks([])
167
![image](images/fft4.jpg)
169
@note As usual, OpenCV functions **cv2.dft()** and **cv2.idft()** are faster than Numpy
170
counterparts. But Numpy functions are more user-friendly. For more details about performance issues,
173
Performance Optimization of DFT
174
===============================
176
Performance of DFT calculation is better for some array size. It is fastest when array size is power
177
of two. The arrays whose size is a product of 2’s, 3’s, and 5’s are also processed quite
178
efficiently. So if you are worried about the performance of your code, you can modify the size of
179
the array to any optimal size (by padding zeros) before finding DFT. For OpenCV, you have to
180
manually pad zeros. But for Numpy, you specify the new size of FFT calculation, and it will
181
automatically pad zeros for you.
183
So how do we find this optimal size ? OpenCV provides a function, **cv2.getOptimalDFTSize()** for
184
this. It is applicable to both **cv2.dft()** and **np.fft.fft2()**. Let's check their performance
185
using IPython magic command %timeit.
187
In [16]: img = cv2.imread('messi5.jpg',0)
188
In [17]: rows,cols = img.shape
189
In [18]: print rows,cols
192
In [19]: nrows = cv2.getOptimalDFTSize(rows)
193
In [20]: ncols = cv2.getOptimalDFTSize(cols)
194
In [21]: print nrows, ncols
197
See, the size (342,548) is modified to (360, 576). Now let's pad it with zeros (for OpenCV) and find
198
their DFT calculation performance. You can do it by creating a new big zero array and copy the data
199
to it, or use **cv2.copyMakeBorder()**.
201
nimg = np.zeros((nrows,ncols))
202
nimg[:rows,:cols] = img
207
bottom = nrows - rows
208
bordertype = cv2.BORDER_CONSTANT #just to avoid line breakup in PDF file
209
nimg = cv2.copyMakeBorder(img,0,bottom,0,right,bordertype, value = 0)
211
Now we calculate the DFT performance comparison of Numpy function:
213
In [22]: %timeit fft1 = np.fft.fft2(img)
214
10 loops, best of 3: 40.9 ms per loop
215
In [23]: %timeit fft2 = np.fft.fft2(img,[nrows,ncols])
216
100 loops, best of 3: 10.4 ms per loop
218
It shows a 4x speedup. Now we will try the same with OpenCV functions.
220
In [24]: %timeit dft1= cv2.dft(np.float32(img),flags=cv2.DFT_COMPLEX_OUTPUT)
221
100 loops, best of 3: 13.5 ms per loop
222
In [27]: %timeit dft2= cv2.dft(np.float32(nimg),flags=cv2.DFT_COMPLEX_OUTPUT)
223
100 loops, best of 3: 3.11 ms per loop
225
It also shows a 4x speed-up. You can also see that OpenCV functions are around 3x faster than Numpy
226
functions. This can be tested for inverse FFT also, and that is left as an exercise for you.
228
Why Laplacian is a High Pass Filter?
229
------------------------------------
231
A similar question was asked in a forum. The question is, why Laplacian is a high pass filter? Why
232
Sobel is a HPF? etc. And the first answer given to it was in terms of Fourier Transform. Just take
233
the fourier transform of Laplacian for some higher size of FFT. Analyze it:
237
from matplotlib import pyplot as plt
239
# simple averaging filter without scaling parameter
240
mean_filter = np.ones((3,3))
242
# creating a guassian filter
243
x = cv2.getGaussianKernel(5,10)
246
# different edge detecting filters
247
# scharr in x-direction
248
scharr = np.array([[-3, 0, 3],
251
# sobel in x direction
252
sobel_x= np.array([[-1, 0, 1],
255
# sobel in y direction
256
sobel_y= np.array([[-1,-2,-1],
260
laplacian=np.array([[0, 1, 0],
264
filters = [mean_filter, gaussian, laplacian, sobel_x, sobel_y, scharr]
265
filter_name = ['mean_filter', 'gaussian','laplacian', 'sobel_x', \
266
'sobel_y', 'scharr_x']
267
fft_filters = [np.fft.fft2(x) for x in filters]
268
fft_shift = [np.fft.fftshift(y) for y in fft_filters]
269
mag_spectrum = [np.log(np.abs(z)+1) for z in fft_shift]
272
plt.subplot(2,3,i+1),plt.imshow(mag_spectrum[i],cmap = 'gray')
273
plt.title(filter_name[i]), plt.xticks([]), plt.yticks([])
279
![image](images/fft5.jpg)
281
From image, you can see what frequency region each kernel blocks, and what region it passes. From
282
that information, we can say why each kernel is a HPF or a LPF
287
-# [An Intuitive Explanation of Fourier
288
Theory](http://cns-alumni.bu.edu/~slehar/fourier/fourier.html) by Steven Lehar
289
2. [Fourier Transform](http://homepages.inf.ed.ac.uk/rbf/HIPR2/fourier.htm) at HIPR
290
3. [What does frequency domain denote in case of images?](http://dsp.stackexchange.com/q/1637/818)